Context Engineering

Use When

• Designing what context an agent or model receives, in what order, and under which token budget.
• Choosing memory scopes, retention rules, compaction behavior, or retrieval boundaries.
• Building or debugging RAG, vector search, grounding, citations, and source attribution.
• Investigating missing context, stale memory, irrelevant retrieval, recitation risk, or prompt-context collisions.

Workflow

1. Identify the decision or output the context must support and the failure mode being addressed.
2. Inventory authoritative sources, memory tiers, retrieval paths, tool outputs, and prompt layers.
3. Define precedence, freshness, provenance, privacy, and token-budget rules before choosing implementation details.
4. Separate durable memory from session state, retrieved evidence, and generated summaries.
5. Test retrieval quality and context assembly with representative queries, adversarial omissions, stale sources, and conflicting evidence.
6. Require citations or source handles where outputs must be auditable.
7. Measure answer quality, retrieval relevance, latency, and cost; document what remains unevaluated.